AI Privacy & Safety Basics: What to Share, What to Keep to Yourself

AI assistants are most useful when you tell them things: your plans, your drafts, your data. That’s also exactly why they deserve five minutes of privacy thinking before the habit forms. This guide gives you a clear mental model of where your words go, a short never-paste list, and a technique that lets you get full value from AI while keeping sensitive details out of it.

No fear-mongering here — the practical risks are manageable with a few habits. But the habits work best if you build them on day one.

Where your chat actually goes

When you send a message to ChatGPT, Claude, or Gemini, three separate things can happen to it. Untangling them dissolves most of the confusion:

1. It’s processed to answer you. Your text goes to the company’s servers, the model reads it, you get a reply. Unavoidable — that’s the service. (The model itself learns nothing permanent from this; it’s frozen.)
2. It’s stored as chat history. So you can revisit conversations. Usually deletable, and most tools offer temporary/incognito chats that skip history.
3. It may be used to train future models. This is the one people worry about, and it varies by product and by your settings. Consumer products differ in their defaults — some train on your chats unless you opt out, others only with opt-in. Business/enterprise tiers almost universally do not train on your data. Action item: open your AI tool’s settings today, find the data/training controls, and set them deliberately. It takes ninety seconds.

The realistic mental model isn’t “everything I type is published” — it’s closer to “treat every chat like an email to a competent company: private-ish, retained somewhere, and out of your hands.” Fine for most things. Wrong for some things. Which brings us to the list.

The never-paste list

Some things don’t belong in any consumer AI chat, regardless of settings:

• Passwords, PINs, recovery codes, API keys. No exceptions, no “just this once.” An AI never needs your password to help you.
• Full government ID numbers, bank account and card numbers — yours or anyone’s.
• Other people’s private information without their consent — a friend’s medical situation, an employee’s salary, a client’s dispute, with names attached. It’s their data, not yours to paste.
• Your employer’s confidential material — customer lists, unreleased financials, proprietary code — unless your company has approved an AI tool for exactly that. Ask. Many companies now provide an approved tool (Copilot, enterprise ChatGPT/Claude/Gemini); use that one for work data, and personal accounts for personal things.
• Anything under NDA, attorney-client privilege, or regulatory protection (patient data, etc.).

Everything else — your drafts, your ideas, your questions, your non-identifying business problems — is the normal, intended use of these tools. Don’t let caution cost you the value.

The anonymization trick (have it both ways)

Here’s the technique that resolves 90% of real-world dilemmas: swap identifying details for placeholders before pasting, and swap them back after. The AI’s help is exactly as good — it never needed the real names.

Instead of:

“My employee Maria Santos in Cebu keeps missing deadlines after her salary dispute in March, here’s her last email…”

paste:

My employee [NAME] keeps missing deadlines after a pay dispute a few
months ago. Here's their last email, lightly edited: [paste with names
and specifics removed]. Help me draft a firm but kind check-in message.

The advice you get back is identical in quality. Rule of thumb: the AI needs the situation, not the identities. Numbers can be rounded, names replaced with roles (“my supplier,” “my sister”), companies with their industry. For recurring use, ask the AI itself to help:

I'm going to paste a document. First, list every piece of personally
identifying information in it so I can decide what to remove. Don't
process the content yet.

Verifying advice that matters

Privacy is one half of AI safety; the other half is acting on wrong answers. The short version (the hallucinations guide is the long version):

• For health, legal, financial, and tax questions: use AI to understand and prepare — explain the jargon, list the questions to ask, draft the letter — but put a qualified human between AI output and real-world action.
• For anything with money on the line, verify load-bearing facts against a primary source first.
• Watch for AI answers about Philippine specifics (laws, government processes, local prices) — thinner training data means more confident guessing. Web-search mode plus official government sites beats memory every time.

The scams to brief your family on

The same technology powering your helpful assistant also powers a new generation of scams. You’re reading an AI guide, so you’ll be fine — brief the people who aren’t:

• Voice cloning. A few seconds of someone’s voice (from a video or voicemail) can produce a convincing clone. The “emergency call” from a relative asking for urgent money now sounds exactly like them. Defense: agree on a family code word for real emergencies, and always call back on the person’s known number.
• Deepfake video and images. Celebrity investment endorsements, fake video calls, fabricated photos. Default skepticism for any get-rich video, no matter whose face is on it.
• AI-polished phishing. The broken-English scam email is extinct; phishing now reads fluently and references real details. The tells are what they’ve always been — urgency, secrecy, requests for credentials or gift cards, mismatched sender addresses — not the grammar anymore.
• Fake AI apps. Copycat “ChatGPT” apps and browser extensions that harvest what you type. Install only official apps from official stores, and be aware that a browser extension that “reads your pages” can read everything you type on them.

One conversation with your parents or kids covering these four patterns is worth more than any software setting on this page.

Kids, students, and honest use

If your household includes students: AI is the best tutor they’ll ever have and the easiest way to not learn anything, depending entirely on how it’s used. The line to teach: AI to understand, your own words to demonstrate. “Explain photosynthesis three ways, then quiz me” builds knowledge; “write my essay” builds nothing and gets caught increasingly often. Same tool, opposite outcomes — and it’s the same line that separates professionals who compound skills from those who quietly deskill.

Your five-minute action list

1. Open your AI tool’s data/training settings and set them on purpose.
2. Memorize the never-paste list (or bookmark this page).
3. Practice the anonymization trick on your next sensitive question.
4. Agree on a family code word for emergency calls.
5. If you use AI for work: find out what your employer has approved, and keep work data inside it.

Do those once and the habits carry you. Privacy handled — now go get the value.

Where to go next

• Never used AI properly yet? Start the path: What Is AI, Actually?
• Ready to get real work done: Prompting Basics and the prompt library
• Catching wrong answers before they cost you: Why AI Makes Things Up